| Anbu Huang | Generative AI & Creative Systems |
Research · Build · Share
Anbu Huang
Researching generative intelligence. Building creative systems.
Research Scientist, Author, AI Enthusiast · THU
Research focus
Read my writingI study Generative Models (Diffusion/Flow-Matching) and VLM Foundation Models, for which you can find more in my blog post.
01 / The generative side
Fidelity. Speed. Control.
I study the “impossible triangle” of high fidelity, fast sampling, and strong controllability—designing training and inference methods (distillation, guidance, solver-aware sampling) that expand the Pareto frontier.
02 / The understanding side
Understanding and creation, together.
I study to build native VLM architectures that achieve a true unification of understanding and generation, enabling AI to perceive, reason, and create within a single, cohesive framework.
Previously: I have also worked on areas including recommender systems, federated learning and AI safety. I have published multiple research papers at AI conferences such as ICLR and AAAI (Publication).
Selected project / Code2X
Genesis
One idea. Three ways to create.
A video creation studio connecting code-driven animation, model-generated storytelling, and hybrid exploration. From precise explanations to expressive visual stories.
Make the abstract
visible.
Describe a concept. Genesis plans the explanation, writes executable Manim code, and renders an animation with aligned narration and subtitles.
- Geometry, equations & controlled motion
- Chinese or English AI narration
- Download video, source code & timeline
From a brief
to a visual story.
Start with a story brief. Plan a sequence, guide characters and locations with visual references, then generate and assemble the individual shots.
- Character, scene & style references
- Optional shot review & refinement
- Visual memory for continuity guidance
Precise by design.
Expressive by nature.
Use code to define structure and motion, then explore model-driven atmosphere and texture. A creative direction for bringing both approaches into one workflow.
- Structure to imagery
- Animation to texture
- Assets to scenes
The hybrid workspace is a UI preview; generation is not enabled yet.
Explore the hybrid previewInside the pipeline — from intent to a finished video
01 / Code-driven animation
- Plan the explanationBuild an outline and storyboard from the topic, duration and visual style.
- Write, voice and renderGenerate shot-level narration, synthesize speech, write Manim code and repair rendering errors.
- Align and exportMatch animation to narration and subtitles. Export MP4, WAV, SRT / VTT, source code and the timeline.
02 / Model-generated stories
- Define the story worldPlan 2–8 shots from a brief. Attach reference images to specific characters, scenes or the overall style.
- Generate and review shotsOptionally inspect sampled frames, revise prompts and select candidates. Continuity guidance tracks visual references and scene state.
- Assemble the storyNormalize the selected clips and join them on a shared timeline, with storyboard and production records available for inspection.
Shot refinement and continuity guidance are independent options. Visual review uses sampled frames; consistency depends on the selected model and references.
The web workspace connects to your own local engine and configured models for generation. Genesis is the studio interface of my Code2X project.
Web project on GitHub